Tech • AI • Robotics
GPT-6 'Sol' Soon! Gemini 4.0 Pro Checkpoint, DeepSeek v4.1 Flash, Nano Banana 2.5, & More! AI NEWS!
9.3/10TL;DR
Google, DeepSeek, and OpenAI are all showing signs of rapid model advances, with new checkpoints, faster open models, and unverified reports linking frontier systems to progress on long-unsolved math problems.
KEY POINTS
Google surfaces a new Gemini Pro checkpoint
Google has internally released a new Gemini Pro checkpoint, the first notable update in that line for some time, and the model ID is now publicly visible. Early tests suggest a substantial jump in capability, fueling speculation that it may be part of the upcoming Gemini 4.0 Pro family.
Early output hints at stronger reasoning and generation
One high-profile test had the new checkpoint generate a detailed SVG peacock in zero-shot mode using about 24,000 tokens over roughly six minutes at high reasoning effort. The result was described as stronger than outputs seen from some rival advanced systems, suggesting improved long-context planning and visual code generation.
Google’s checkpoint trail points to active iteration
The previous project checkpoint was reportedly labeled Argon 160 and had also been tested under the Gemini 3.8 Flash name in public arenas. That pattern indicates Google is continuing to refine multiple Gemini variants behind the scenes before a broader Pro-series launch.
A new “Nanobanana” image model may also be nearing release
A mysterious Google image model using the codename Spicy Mayo has appeared in model arenas and may be tied to a new NanoBanana 2.5 or a lighter update. In one comparison against GPT Image 2.5 on a Minecraft camel prompt, Google’s model produced the more visually appealing image, while OpenAI’s was judged more accurate to the game interface.
DeepSeek launches version 4.1 Flash
DeepSeek has officially released DeepSeek 4.1 Flash, positioning it as a smarter, faster, and more efficient model with native vision support. The model uses a 552 billion parameter architecture, but activates only about 8 billion parameters for input and 16 billion for output, a design aimed at delivering stronger results with less compute.
Benchmarks place DeepSeek among top-performing models
Reported testing ranks DeepSeek 4.1 Flash seventh overall across a broad benchmark set, with especially strong results in coding and agentic workflows. On Terminal Bench 2.1, it reportedly scored 90.6, placing it ahead of several larger-name competitors, including some flagship-class models and even DeepSeek 4 Pro on that measure.
Speed and pricing are central to DeepSeek’s pitch
The model is said to run at roughly 300 to 400 tokens per second, making it unusually fast for its capability level. Pricing was reported at $0.30 per million input tokens and $1.20 per million output tokens, with lower off-peak rates, strengthening its case as a low-cost alternative to premium systems.
DeepSeek trades cost for quality in some visual coding tasks
In a six-task 3D geometry comparison against Gemini 3.8 Flash, Google’s model was more efficient and cheaper, costing about $0.11 versus $0.27 for DeepSeek. But the qualitative outputs from DeepSeek 4.1 Flash were judged better on several tasks, showing a tradeoff between cost efficiency and final visual quality.
OpenAI’s next releases are starting to appear
A model referred to as GPT-6 Sol has reportedly been spotted in public-facing systems after rumors of internal testing. Early signs point to a faster model with improvements over GPT-5.6 Sol, possibly part of a tiered rollout strategy ahead of a broader release window later in September.
Math rumors around OpenAI and Anthropic are intensifying
Separate reports claim OpenAI may be nearing verification of the Hodge conjecture and making progress on the Birch and Swinnerton-Dyer conjecture, both among the most famous unsolved problems in mathematics. Those claims remain unverified, but they follow earlier reporting tying an OpenAI Bell checkpoint to solving a decade-old mathematics problem within days, and broader suggestions that both OpenAI and Anthropic are training systems well beyond their current public models.
OpenAI is also expanding product features
ChatGPT Voice is reportedly gaining access to newer internal model variants including GPT-5.16 Sol and GPT-6 Astra, a move expected to improve intelligence and freshness of responses. OpenAI has also introduced a new data agent in ChatGPT Work, allowing companies to connect internal data sources and turn them into answers, dashboards, and actions.
CONCLUSION
The latest developments point to an AI market moving on several fronts at once: stronger frontier checkpoints, cheaper high-performance open models, and early signs that advanced systems may be entering serious scientific territory. If even part of the mathematics reporting is confirmed, the competitive race between Google, DeepSeek, OpenAI, and Anthropic will take on far broader significance than product launches alone.
Explain this